A latent counterfactual disentanglement framework for implicit hate speech detection

Implicit hate speech often conceals hateful intent through semantic camouflage, such as sarcasm and metaphor, causing core toxicity features to remain hidden beneath the surface text. Existing models tend to jointly encode core toxicity and camouflage style during feature extraction, thereby leading to spurious decision shortcuts dominated by camouflage-style cues. To alleviate this problem, this paper approaches implicit hate speech detection from the perspective of camouflage-style removal and proposes a latent counterfactual disentanglement (LCD) framework for implicit hate speech detection. Specifically, LCD decomposes textual latent representations into a core toxicity subspace and a camouflage style subspace through causal feature disentanglement. It further constructs counterfactually augmented representations using a feature-level multiple counterfactual intervention mechanism, thereby reducing the model’s dependence on camouflage style. Meanwhile, a latent semantic reconstruction module is introduced to mitigate the risk of information loss during disentanglement. Experimental results on three challenging public benchmark datasets, namely IHC, SBIC, and DYNA, show that LCD achieves Macro-F1 scores of 79.20%, 91.09%, and 83.43%, respectively. Further experiments and visualization analyses demonstrate that LCD can reduce the leakage of label-related information into the camouflage style branch and improve prediction consistency under style perturbations.